KODA · AI Open House
Start clock
The claim: one automation, measured in hours

One autonomous action. Here is the hour a day it gave me back.

My system runs a lot of things. This is just one of them: a single autonomous action that turns every meeting and every commitment into a tracked list. Here is what that one change did to my daily workload.
Before
1 to 2 hrs/day
Filtering email and digging back through my notes to be sure I had followed up on everything. Roughly 5 to 10 hours a week.
→
After
~10 min/day
A quick review of a list that is already complete. The time comes back, every day.
→
The forecast
~300 hrs/year
About 25 hours back every month, and roughly 7 full work weeks a year. From this one automation, and it compounds as I add more.
And I don't miss things anymore. If I say I'll send something Thursday on a call, or someone asks me to do something, it's on my list before I've hung up.
Same me. More capacity. Nothing dropped.
The workflow, tool by tool

Every meeting becomes tracked action items, while I'm onto the next thing.

🎤
Plaud
I record the meeting on my Plaud. It transcribes itself.
→
⚡
Zapier
Catches the finished transcript automatically.
→
📦
Dropbox
Drops it into an inbox folder I never open.
→
🤖
Claude
Pulls it, extracts decisions, owners, and every commitment.
→
📋
Action list
Items land in my vault, tagged to the meeting.
→
📊
Claude
Rolls it all up into the dashboard you just saw.
I record. Everything after the microphone runs on its own.
The build, the technical part

The stack, and the parts that actually mattered.

The stack
Claude Code
the agent that reads and acts
Markdown + Git
the brain: my real context, versioned
Python skills
deterministic engines (rollup, brief)
Zapier + Dropbox
capture and file movement
Google, plus more
Calendar and Gmail connectors
Scheduled agents
cron: they run themselves
What broke, and what I'd tell you
Permissions were the whole fight
An unattended agent stalls on approval prompts. The fix is one single allowlisted command per job. A compound cd && python can never be pre approved. That one detail is the difference between "runs itself" and "nags you."
Sanitize, then firewall the sensitive data
This is the part almost nobody budgets time for, and it is what makes it safe to run on real data. The work is in the plumbing:
Scrub first: strip identifiers (names, record numbers, dollar figures) on the machine, before anything is sent up for processing, not after.
Firewall at rest: anything sensitive stays git ignored and local only, never pushed or synced out.
Pointers, not payloads: the calendar and every outside surface get a link, never the content.
Deterministic beats clever
The generator is a fixed output Python skill, not a free form prompt. Identical every run, so it is safe to schedule and safe to demo.
Everything links back
Each item deep links to the call it came from, so "where did this come from?" is one click.
Zoom out

This is one gear. Here is the whole machine.

6
agents on a clock
11
connectors wired in
71
skills built in the vault

Agents that run themselves

Meeting to filed notes and action items (nightly)
Action rollup: reconcile done, rebuild the list (nightly)
Monday command center (what you saw)
Weekly AI intel sweep: what shipped, what we can use (Sundays)
Nonprofit outreach pipeline: research and draft (daily)
Monthly board report draft (third Monday)

Connected to everything I use

✉️ Gmail
📅 Google Calendar
📁 Google Drive
📦 Dropbox
⚡ Zapier (+9,000 apps)
🎨 Canva
👷 Lovable
📝 Microsoft Word
📄 PDF tools
🌐 Chrome control
🖥️ Mac desktop control
Same pattern every time: a brain of real context, plus Claude, plus a deterministic skill, equals a thing that runs itself.
71 skills, built and stored in the vault (the library the agents draw from)
The leave behind, the bigger idea

You don't need my job to build this. You need your context.

1

Meet yourself on the curve

Chat, then copilots, then agents. Wherever you are is a valid start. Build from your habits, not from zero.

2

Build the brain

A private, organized layer of your real company context: decisions, policies, voice. Not what is floating on the internet.

3

Feed it relentlessly

Meetings, notes, finished work. If it isn't written down it doesn't exist, and every addition makes every future answer better.

4

Model agents on you

So they don't just fetch. They act and react the way you would, on your standards, for your business.

⭐
The biggest takeaway: it works with any model. The brain is plain text, so it is not tied to Claude. Switch to another AI, or run several at once, and your brain and every bit of context stay put. As AI keeps changing, you never rebuild and you never lose your history.
📁 Inside the starter brain (what I'm leaving with the facilitator)
A working brain you clone, point at your business, and run. Everything it needs to answer like your best employee on day one.
▸
CLAUDE.md how the assistant should work in your business
▸
Company Overview what you do, how you win and lose work
▸
Team and Roles who owns what
▸
Policies and Procedures your rules, so answers match them
▸
Brand Voice so drafts sound like you, not a robot
▸
Glossary your words, defined once
▸
Meeting Notes and Action Items the running record
▸
Starter skills meeting to actions, weekly rollup, grounded Q&A
▸
Your files stay put connected to your existing Drive or Dropbox
Clone it, point it at your business, and it runs. I'm leaving the starter brain and the recipe with the facilitator.
AI Open House · Session 001 Arrow keys or click a tab to move · click the clock to time yourself